Conditioning on Colliders: Selection Bias
What you'll learn: Why controlling for a collider creates spurious associations and opens the door to selection bias.
The Collider Problem
You already know from your DAG studies that a collider is a variable caused by two or more other variables (arrows point into it). Here's the critical rule: conditioning on a collider opens a non-causal path between its parent variables, creating a spurious association where none existed before.
Think of it like this: Imagine talent and attractiveness both independently cause someone to become a celebrity. If you study only celebrities (conditioning on the collider "celebrity status"), you'll falsely observe that among celebrities, talented people seem less attractive and vice versa. Why? Because to make it into your sample without high talent, someone likely needed high attractiveness — and the reverse. This is selection bias.
Common Examples
Job performance studies: Both competence and charisma can lead to promotion. If you study only promoted employees, you might wrongly conclude competence and charisma are negatively correlated — because those weak in one dimension needed to be strong in the other to get promoted.
Hospital studies: Disease severity and access to care both lead to hospitalization. Analyzing only hospitalized patients can create false relationships between severity and other factors.
Dating app data: Attractiveness and personality both lead to getting matches. Analyzing only matched pairs can distort the true relationship between these traits.
Why This Matters
Unlike confounders (which you should condition on), colliders should be avoided in your adjustment set. Conditioning on them manufactures bias where there was none. This is a major source of misleading conclusions in observational data analysis.
Key Takeaway: Conditioning on a collider opens a non-causal backdoor path, inducing spurious correlations through selection bias — avoid adjusting for colliders when estimating causal effects.